Liver Imaging Reporting and Data System Contrast-Enhanced US Nonradiation Treatment Response Assessment Version 2024.

The American College of Radiology Liver Imaging Reporting and Data System (LI-RADS) standardizes the imaging technique, reporting lexicon, disease categorization, and management for patients with or at risk for hepatocellular carcinoma (HCC). LI-RADS encompasses HCC surveillance with US; HCC diagnosis with CT, MRI, or contrast-enhanced US (CEUS); and treatment response assessment (TRA) with CT or MRI. LI-RADS was recently expanded to include CEUS TRA after nonradiation locoregional therapy or surgical resection. This report provides an overview of LI-RADS CEUS Nonradiation TRA v2024, including a lexicon of imaging findings, techniques, and imaging criteria for posttreatment tumor viability assessment. LI-RADS CEUS Nonradiation TRA v2024 takes into consideration differences in the CEUS appearance of viable tumor and posttreatment changes within and in close proximity to a treated lesion. Due to the high sensitivity of CEUS to vascular flow, posttreatment reactive changes commonly manifest as areas of abnormal perilesional enhancement without washout, especially in the first 3 months after treatment. To improve the accuracy of CEUS for nonradiation TRA, different diagnostic criteria are used to evaluate tumor viability within and outside of the treated lesion margin. Broader criteria for intralesional enhancement increase sensitivity for tumor viability detection. Stricter criteria for perilesional enhancement limit miscategorization of posttreatment reactive changes as viable tumor. Finally, the TRA algorithm reconciles intralesional and perilesional tumor viability assessment and assigns a single LI-RADS treatment response (LR-TR) category: LR-TR nonviable, LR-TR equivocal, or LR-TR viable.

Breast Imaging Reporting and Data System (BI-RADS®): a success history and particularities of its use in Brazil.

BI-RADS® is a standardization system for breast imaging reports and results created by the American College of Radiology to initially address the lack of uniformity in mammography reporting. The system consists of a lexicon of descriptors, a reporting structure with final categories and recommended management, and a structure for data collection and auditing. It is accepted worldwide by all specialties involved in the care of breast diseases. Its implementation is related to the Mammography Quality Standards Act initiative in the United States (1992) and breast cancer screening. After its initial creation in 1993, four additional editions were published in 1995, 1998, 2003 and 2013. It is adopted in several countries around the world and has been translated into 6 languages. Successful breast cancer screening programs in high-income countries can be attributed in part to the widespread use of BI-RADS®. This success led to the development of similar classification systems for other organs (e.g., lung, liver, thyroid, ovaries, colon). In 1998, the structured report model was adopted in Brazil. This article highlights the pioneering and successful role of BI-RADS®, created by ACR 30 years ago, on the eve of publishing its sixth edition, which has evolved into a comprehensive quality assurance tool for multiple imaging modalities. And, especially, it contextualizes the importance of recognizing how we are using BI-RADS® in Brazil, from its implementation to the present day, with a focus on breast cancer screening.

Transforming Tele-Ophthalmology: Utilizing Cloud Computing for Remote Eye Care.

BACKGROUND: Tele-ophthalmology is gaining recognition for its role in improving eye care accessibility via cloud-based solutions. The Google Cloud Platform (GCP) Healthcare API enables secure and efficient management of medical image data such as high-resolution ophthalmic images. OBJECTIVES: This study investigates cloud-based solutions' effectiveness in tele-ophthalmology, with a focus on GCP's role in data management, annotation, and integration for a novel imaging device. METHODS: Leveraging the Integrating the Healthcare Enterprise (IHE) Eye Care profile, the cloud platform was utilized as a PACS and integrated with the Open Health Imaging Foundation (OHIF) Viewer for image display and annotation capabilities for ophthalmic images. RESULTS: The setup of a GCP DICOM storage and the OHIF Viewer facilitated remote image data analytics. Prolonged loading times and relatively large individual image file sizes indicated system challenges. CONCLUSION: Cloud platforms have the potential to ease distributed data analytics, as needed for efficient tele-ophthalmology scenarios in research and clinical practice, by providing scalable and secure image management solutions.

Clinical utility, reproducibility, and radiologist acceptance of ILD-RADS.

PURPOSE: To investigate the clinical utility, reproducibility, and radiologists' acceptance of the Interstitial Lung Disease Imaging-Reporting and Data System (ILD-RADS). METHOD: In this single-institutional retrospective study, three radiologists independently reviewed the chest high-resolution CT (HRCT) scans of 111 consecutive patients diagnosed with ILDs. They assessed the HRCT pulmonary features using the ILD-RADS template and assigned an ILD-RADS category (1-4) to each scan based on the identified imaging pattern. Patients were classified into idiopathic pulmonary fibrosis (IPF) (n = 14) and non-IPF ILD (n = 97) groups based on clinical diagnoses determined by multidisciplinary discussion. Association between ILD-RADS categories and clinical diagnoses was assessed using the Chi-square test for trend. Reproducibility was evaluated using kappa (k) scores, and radiologists' acceptance of the ILD-RADS was evaluated with a questionnaire. RESULTS: We found a significant association between the ILD-RADS categories and patients' clinical diagnoses (P ≤ 0.0001) for the three readers, with a trend toward increased assignment of ILD-RADS-1 to IPF patients (50 %-57.1 %), and ILD-RADS-4 to non-IPF patients (46.4 %-49.5 %). The ILD-RADS categories showed excellent intra-reader agreement (k = 0.873) and moderate inter-reader agreement (k = 0.440). ILD-RADS-1 and -4 categories showed the highest inter-reader agreement (k = 0.681 and 0.481, respectively). Radiologists gave a positive response to using the ILD-RADS in daily practice. CONCLUSIONS: The clinical utility of the ILD-RADS was demonstrated by the significant association between the ILD-RADS categories and patients' clinical diagnoses, particularly the ILD-RADS-1 and -4 categories. Excellent intra-reader and moderate inter-reader reproducibility was observed. ILD-RADS has the potential to be widely accepted for standardized HRCT reporting among radiologists.

BI-RADS Category Assignments by GPT-3.5, GPT-4, and Google Bard: A Multilanguage Study.

Background The performance of publicly available large language models (LLMs) remains unclear for complex clinical tasks. Purpose To evaluate the agreement between human readers and LLMs for Breast Imaging Reporting and Data System (BI-RADS) categories assigned based on breast imaging reports written in three languages and to assess the impact of discordant category assignments on clinical management. Materials and Methods This retrospective study included reports for women who underwent MRI, mammography, and/or US for breast cancer screening or diagnostic purposes at three referral centers. Reports with findings categorized as BI-RADS 1-5 and written in Italian, English, or Dutch were collected between January 2000 and October 2023. Board-certified breast radiologists and the LLMs GPT-3.5 and GPT-4 (OpenAI) and Bard, now called Gemini (Google), assigned BI-RADS categories using only the findings described by the original radiologists. Agreement between human readers and LLMs for BI-RADS categories was assessed using the Gwet agreement coefficient (AC1 value). Frequencies were calculated for changes in BI-RADS category assignments that would affect clinical management (ie, BI-RADS 0 vs BI-RADS 1 or 2 vs BI-RADS 3 vs BI-RADS 4 or 5) and compared using the McNemar test. Results Across 2400 reports, agreement between the original and reviewing radiologists was almost perfect (AC1 = 0.91), while agreement between the original radiologists and GPT-4, GPT-3.5, and Bard was moderate (AC1 = 0.52, 0.48, and 0.42, respectively). Across human readers and LLMs, differences were observed in the frequency of BI-RADS category upgrades or downgrades that would result in changed clinical management (118 of 2400 [4.9%] for human readers, 611 of 2400 [25.5%] for Bard, 573 of 2400 [23.9%] for GPT-3.5, and 435 of 2400 [18.1%] for GPT-4; P < .001) and that would negatively impact clinical management (37 of 2400 [1.5%] for human readers, 435 of 2400 [18.1%] for Bard, 344 of 2400 [14.3%] for GPT-3.5, and 255 of 2400 [10.6%] for GPT-4; P < .001). Conclusion LLMs achieved moderate agreement with human reader-assigned BI-RADS categories across reports written in three languages but also yielded a high percentage of discordant BI-RADS categories that would negatively impact clinical management. © RSNA, 2024 Supplemental material is available for this article.

Evaluation of large language models performance against humans for summarizing MRI knee radiology reports: A feasibility study.

OBJECTIVES: This study addresses the critical need for accurate summarization in radiology by comparing various Large Language Model (LLM)-based approaches for automatic summary generation. With the increasing volume of patient information, accurately and concisely conveying radiological findings becomes crucial for effective clinical decision-making. Minor inaccuracies in summaries can lead to significant consequences, highlighting the need for reliable automated summarization tools. METHODS: We employed two language models - Text-to-Text Transfer Transformer (T5) and Bidirectional and Auto-Regressive Transformers (BART) - in both fine-tuned and zero-shot learning scenarios and compared them with a Recurrent Neural Network (RNN). Additionally, we conducted a comparative analysis of 100 MRI report summaries, using expert human judgment and criteria such as coherence, relevance, fluency, and consistency, to evaluate the models against the original radiologist summaries. To facilitate this, we compiled a dataset of 15,508 retrospective knee Magnetic Resonance Imaging (MRI) reports from our Radiology Information System (RIS), focusing on the findings section to predict the radiologist's summary. RESULTS: The fine-tuned models outperform the neural network and show superior performance in the zero-shot variant. Specifically, the T5 model achieved a Rouge-L score of 0.638. Based on the radiologist readers' study, the summaries produced by this model were found to be very similar to those produced by a radiologist, with about 70% similarity in fluency and consistency between the T5-generated summaries and the original ones. CONCLUSIONS: Technological advances, especially in NLP and LLM, hold great promise for improving and streamlining the summarization of radiological findings, thus providing valuable assistance to radiologists in their work.

Transforming free-text radiology reports into structured reports using ChatGPT: A study on thyroid ultrasonography.

PURPOSE: The importance of structured radiology reports has been fully recognized, as they facilitate efficient data extraction and promote collaboration among healthcare professionals. Our purpose is to assess the accuracy and reproducibility of ChatGPT, a large language model, in generating structured thyroid ultrasound reports. METHODS: This is a retrospective study that includes 184 nodules in 136 thyroid ultrasound reports from 136 patients. ChatGPT-3.5 and ChatGPT-4.0 were used to structure the reports based on ACR-TIRADS guidelines. Two radiologists evaluated the responses for quality, nodule categorization accuracy, and management recommendations. Each text was submitted twice to assess the consistency of the nodule classification and management recommendations. RESULTS: On 136 ultrasound reports from 136 patients (mean age, 52 years ± 12 [SD]; 61 male), ChatGPT-3.5 generated 202 satisfactory structured reports, while ChatGPT-4.0 only produced 69 satisfactory structured reports (74.3 % vs. 25.4 %, odds ratio (OR) = 8.490, 95 %CI: 5.775-12.481, p < 0.001). ChatGPT-4.0 outperformed ChatGPT-3.5 in categorizing thyroid nodules, with an accuracy of 69.3 % compared to 34.5 % (OR = 4.282, 95 %CI: 3.145-5.831, p < 0.001). ChatGPT-4.0 also provided more comprehensive or correct management recommendations than ChatGPT-3.5 (OR = 1.791, 95 %CI: 1.297-2.473, p < 0.001). Finally, ChatGPT-4.0 exhibits higher consistency in categorizing nodules compared to ChatGPT-3.5 (ICC = 0.732 vs. ICC = 0.429), and both exhibited moderate consistency in management recommendations (ICC = 0.549 vs ICC = 0.575). CONCLUSIONS: Our study demonstrates the potential of ChatGPT in transforming free-text thyroid ultrasound reports into structured formats. ChatGPT-3.5 excels in generating structured reports, while ChatGPT-4.0 shows superior accuracy in nodule categorization and management recommendations.

IODeep: An IOD for the introduction of deep learning in the DICOM standard.

BACKGROUND AND OBJECTIVE: In recent years, Artificial Intelligence (AI) and in particular Deep Neural Networks (DNN) became a relevant research topic in biomedical image segmentation due to the availability of more and more data sets along with the establishment of well known competitions. Despite the popularity of DNN based segmentation on the research side, these techniques are almost unused in the daily clinical practice even if they could support effectively the physician during the diagnostic process. Apart from the issues related to the explainability of the predictions of a neural model, such systems are not integrated in the diagnostic workflow, and a standardization of their use is needed to achieve this goal. METHODS: This paper presents IODeep a new DICOM Information Object Definition (IOD) aimed at storing both the weights and the architecture of a DNN already trained on a particular image dataset that is labeled as regards the acquisition modality, the anatomical region, and the disease under investigation. RESULTS: The IOD architecture is presented along with a DNN selection algorithm from the PACS server based on the labels outlined above, and a simple PACS viewer purposely designed for demonstrating the effectiveness of the DICOM integration, while no modifications are required on the PACS server side. Also a service based architecture in support of the entire workflow has been implemented. CONCLUSION: IODeep ensures full integration of a trained AI model in a DICOM infrastructure, and it is also enables a scenario where a trained model can be either fine-tuned with hospital data or trained in a federated learning scheme shared by different hospitals. In this way AI models can be tailored to the real data produced by a Radiology ward thus improving the physician decision making process. Source code is freely available at https://github.com/CHILab1/IODeep.git.

Benign thyroid nodules classified as ACR TI-RADS 4 or 5: Imaging and histological features.

BACKGROUND: American College of Radiology Thyroid Imaging Reporting and Data System (ACR TI-RADS) being most widely applied in clinical practice, there is an overlap in US imaging manifestations between benign and malignant thyroid nodules. OBJECTIVES: To analyze the imaging and histological characteristics of pathological benign thyroid nodules categorized as American College of Radiology Thyroid Imaging Reporting and Data System (ACR TI-RADS) 4 or 5, and to explore the correlation between the suspicious sonographic signs resulting in the misdiagnoses and the histopathological features. MATERIALS AND METHODS: Overall, 227 benign thyroid nodules (215 patients) in ACR TI-RADS 4 or 5 sampled through surgical excision were analyzed between December 2020 and August 2022. We retrospectively reread the ultrasound (US) images of the pathological discordant cases, after which we performed a systematic analysis focusing on the histopathological characteristics of thyroid lesions and recorded the findings. Qualitative US features and pathological significance of the thyroid nodules were analyzed using the chi-square and Fisher's exact tests. RESULTS: The pathological type of 227 thyroid nodules (n = 103 in ACR TI-RADS 4 and n = 124 in ACR TI-RADS 5) was nodular goiter together with other histopathological features, namely, fibrosis (n = 103, 45.4 %), calcification (n = 70, 30.8 %), adenomatous hyperplasia (n = 31, 13.7 %), follicular epithelial hyperplasia (n = 23, 10.1 %), Hashimoto's thyroiditis (n = 18, 7.9 %), and cystic degeneration (n = 16, 7.1 %). Fibrosis was the most common histopathological feature in both ACR TI-RADS 4 (n = 42, 40.8 %) and 5 (n = 61, 49.2 %) categories of benign thyroid nodules. Thyroid nodules with fibrosis demonstrated sonographic features of "taller than wide" (p < 0.05), while lesions with follicular epithelial hyperplasia were likely to be detected with irregular and/or lobulated margins and very hypoechoic on US (p < 0.05 for both). CONCLUSION: Benign thyroid nodules with histopathological findings such as fibrosis are associated with suspicious US features, which may give inappropriately higher TIRADS stratification.

Performance of VI-RADS in predicting muscle-invasive bladder cancer after transurethral resection: a single center retrospective analysis.

PURPOSE: To assess VIRADS performance and inter-reader agreement for detecting muscle-invasive bladder cancer (MIBC) following transurethral resection of bladder tumor (TURBT). METHODS: An IRB-approved, HIPAA-compliant, retrospective study from 2016 to 2020 included patients with bladder urothelial carcinoma who underwent MRI after TURBT, and cystectomy within 3 months without post-MRI treatments. Three radiologists blinded to pathology results independently reviewed MR images and assigned a VI-RADS score. Sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV) and accuracy of VI-RADS were assessed for diagnosing MIBC using VI-RADS scores ≥ 3 and ≥ 4. Inter-reader agreement was assessed using Gwet's agreement coefficient (AC) and percent agreement. RESULTS: The cohort consisted of 70 patients (mean age, 68 years ± 11 [SD]; range 39-85; 58 men) and included 32/70 (46%) with MIBC at cystectomy. ROC analysis revealed an AUC ranging from 0.67 to 0.77 and no pairwise statistical difference between readers (p-values, 0.06, 0.08, 0.97). Percent sensitivity, specificity, PPV, NPV and accuracy for diagnosing MIBC for the three readers ranged from 81.3-93.8, 36.8-55.3, 55.6-60.5, 77.3-87.5, and 62.9-67.1 respectively for VI-RADS score ≥ 3, and 78.1-81.3, 47.4-68.4, 55.6-67.6, 72.0-78.8 and 61.4-72.9 respectively for VI-RADS score ≥ 4. Gwet's AC was 0.63 [95% confidence interval (CI): 0.49,0.78] for VI-RADS score ≥ 3 with 79% agreement [95% CI 72,87] and 0.54 [95%CI 0.38,0.70] for VI-RADS score ≥ 4 with 76% agreement [95% CI 69,84]. VIRADS performance was not statistically different among 31/70 (44%) patients who received treatments prior to MRI (p ≥ 0.16). CONCLUSION: VI-RADS had moderate sensitivity and accuracy but low specificity for detection of MIBC following TURBT, with moderate inter-reader agreement.

Memory Guided Transformer With Spatio-Semantic Visual Extractor for Medical Report Generation.

Medicalimaging-based report writing for effective diagnosis in radiology is time-consuming and can be error-prone by inexperienced radiologists. Automatic reporting helps radiologists avoid missed diagnoses and saves valuable time. Recently, transformer-based medical report generation has become prominent in capturing long-term dependencies of sequential data with its attention mechanism. Nevertheless, input features obtained from traditional visual extractor of conventional transformers do not capture spatial and semantic information of an image. So, the transformer is unable to capture fine-grained details and may not produce detailed descriptive reports of radiology images. Therefore, we propose a spatio-semantic visual extractor (SSVE) to capture multi-scale spatial and semantic information from radiology images. Here, we incorporate two types of networks in ResNet 101 backbone architecture, i.e. (i) deformable network at the intermediate layer of ResNet 101 that utilizes deformable convolutions in order to obtain spatially invariant features, and (ii) semantic network at the final layer of backbone architecture which uses dilated convolutions to extract rich multi-scale semantic information. Further, these network representations are fused to encode fine-grained details of radiology images. The performance of our proposed model outperforms existing works on two radiology report datasets, i.e., IU X-ray and MIMIC-CXR.

[Construction of a Temporary Radiological Image Viewing Network for Large-scale Disasters].

The Ministry of Health, Labor and Welfare mandated the creation of the business continuity plan (BCP) for disaster key hospitals on March 31, 2017. Supposing the hospital information system (HIS) failure occurred, the picture archiving and communication system (PACS) also suffers obstacles, we assumed building a new network was necessary for radiological examination images. The purpose of this study was to investigate whether building a new network for radiological examination images is necessary in an emergency. Using wireless fidelity (Wi-Fi), the new network consisting of one image server and two tablet terminals A and B was constructed. The study measured the portable image transfer time for various stages of the network. The results were as follows: Transfer time from the mobile X-ray unit to the image server was 4.12±0.86 s, that from the image server to the tablet device A was 5.14±0.71 s, and that from the image server to the tablet device B was 7.32±1.66 s. Therefore, the new network configuration can provide a reliable means of accessing radiological images during emergency situations when the HIS and PACS may experience obstacles or failures.

Systematic training of LI-RADS CT v2018 improves interobserver agreements and performances in LR categorization for focal liver lesions.

AIM: To retrospectively explored whether systematic training in the use of Liver Imaging Reporting and Data System (LI-RADS) v2018 on computed tomography (CT) can improve the interobserver agreements and performances in LR categorization for focal liver lesions (FLLs) among different radiologists. MATERIALS AND METHODS: A total of 18 visiting radiologists and the liver multiphase CT images of 70 hepatic observations in 63 patients at high risk of HCC were included in this study. The LI-RADS v2018 training procedure included three thematic lectures, with an interval of 1 month. After each seminar, the radiologists had 1 month to adopt the algorithm into their daily work. The interobserver agreements and performances in LR categorization for FLLs among the radiologists before and after training were compared. RESULTS: After training, the interobserver agreements in classifying the LR categories for all radiologists were significantly increased for most LR categories (P < 0.001), except for LR-1 (P = 0.053). After systematic training, the areas under the curve (AUCs) for LR categorization performance for all participants were significantly increased for most LR categories (P < 0.001), except for LR-1 (P = 0.062). CONCLUSION: Systematic training in the use of the LI-RADS can improve the interobserver agreements and performances in LR categorization for FLLs among radiologists with different levels of experience.

General-Purpose Large Language Models Versus a Domain-Specific Natural Language Processing Tool for Label Extraction From Chest Radiograph Reports.

GPT-4 outperformed a radiology domain-specific natural language processing model in classifying imaging findings from chest radiograph reports, both with and without predefined labels. Prompt engineering for context further improved performance. The findings indicate a role for large language models to accelerate artificial intelligence model development in radiology by automating data annotation.

The Utility of ACR TI-RADS in Predicting False-Negative Fine Needle Aspiration for Thyroid Cancer.

BACKGROUND: Thyroid nodule fine needle aspiration (FNA) biopsies are associated with a low false-negative rate. There is limited data regarding the predictive value of American College of Radiology Thyroid Imaging Reporting and Data System for false-negative FNA. METHODS: This single-center retrospective study evaluated 119 patients who underwent thyroidectomy. The association of TR category, along with other clinical variables, with false-negative FNA was evaluated. RESULTS: The overall false-negative rate of FNA was 10.8% (n = 9). False-negative FNAs were associated with younger age (mean 42 years vs 50.6 years, P = .04), larger nodule size (mean 4.4 cm vs 3.2 cm, P = .03), and a lower TR category (median 3 v 4, P = .01). DISCUSSION: Lower TR category, younger age, and larger nodule size were associated with false-negative FNA of thyroid nodules. These findings should be taken into context when counseling patients with thyroid nodules who have a benign FNA.

Katherine P. Andriole - A trailblazer in medical informatics and radiology.

In a traditionally male-dominated field, the journey of Dr. Andriole represents a pioneering path in the realms of radiology and medical imaging informatics. Her career has not only reshaped the landscape of radiology but also championed diversity, equity, and inclusion in healthcare technology. Through a comprehensive exploration of Dr. Andriole's career trajectory, we navigate her transition from analog to digital radiology, her influential role in pioneering picture archiving communication systems (PACS), and her dedication to mentorship and education in the field. Dr. Andriole's journey underscores the growing influence of women in radiology and informatics, exemplified by her Gold Medal accolades from esteemed organizations. Dr. Andriole's career serves as a beacon for aspiring radiologists and informaticians, emphasizing the significance of passion, mentorship, and collaborative teamwork in advancing the fields of radiology and informatics.

From omission to excellence: Reducing the percentage of incomplete trainee reports by ensuring documentation of comparison studies.

The inclusion of comparison studies within radiology reports is an important, standard practice. Despite this, we identified that after-hours preliminary reports rendered by trainees within our institution often omitted reference to comparison studies for pediatric inpatient portable radiographs. We addressed this issue through a quality improvement project targeting pediatric radiographs. Key interventions included modifying the structured reports by removing default text in the comparison field, designating the comparison field as mandatory, and restructuring the report templates to remove extraneous information. We also initiated a targeted educational campaign. 392 reports before and 267 reports after intervention (total 732 reports) were evaluated to determine the number of reports lacking comparison information when comparisons were available. Following the interventions, there was a statistically significant decrease in incomplete reports from 12.5% to 6%. This project highlights the success of utilizing structured reporting to improve the quality of trainee reports.

Generative Large Language Models for Detection of Speech Recognition Errors in Radiology Reports.

This study evaluated the ability of generative large language models (LLMs) to detect speech recognition errors in radiology reports. A dataset of 3233 CT and MRI reports was assessed by radiologists for speech recognition errors. Errors were categorized as clinically significant or not clinically significant. Performances of five generative LLMs-GPT-3.5-turbo, GPT-4, text-davinci-003, Llama-v2-70B-chat, and Bard-were compared in detecting these errors, using manual error detection as the reference standard. Prompt engineering was used to optimize model performance. GPT-4 demonstrated high accuracy in detecting clinically significant errors (precision, 76.9%; recall, 100%; F1 score, 86.9%) and not clinically significant errors (precision, 93.9%; recall, 94.7%; F1 score, 94.3%). Text-davinci-003 achieved F1 scores of 72% and 46.6% for clinically significant and not clinically significant errors, respectively. GPT-3.5-turbo obtained 59.1% and 32.2% F1 scores, while Llama-v2-70B-chat scored 72.8% and 47.7%. Bard showed the lowest accuracy, with F1 scores of 47.5% and 20.9%. GPT-4 effectively identified challenging errors of nonsense phrases and internally inconsistent statements. Longer reports, resident dictation, and overnight shifts were associated with higher error rates. In conclusion, advanced generative LLMs show potential for automatic detection of speech recognition errors in radiology reports. Keywords: CT, Large Language Model, Machine Learning, MRI, Natural Language Processing, Radiology Reports, Speech, Unsupervised Learning Supplemental material is available for this article.